AI Automated the Interview. It Shouldn't Automate the Decision.
TL;DR: AI-moderated interview platforms let companies collect qualitative research from hundreds of customers in days instead of months. Anthropic, Microsoft, and Sweetgreen all use this approach today. At each company, a research lead still decides what the data means and what gets built. The platform only collects it.
What does AI automate in customer research today?
The part that used to be the bottleneck. I spent years running traditional qualitative studies before Monadux — recruiting participants, scheduling calls, moderating one conversation at a time — and the ceiling never moved: a research team could talk to a few dozen people before a deadline forced a synthesis. AI-moderated interview platforms remove that ceiling by running adaptive conversations with hundreds of participants at once, in whatever format (video, voice, or text) a respondent prefers.
A Harvard Business Review piece on the shift, co-authored by Wharton's Stefano Puntoni and Columbia Business School's Olivier Toubia, describes AI-moderated interviews compressing research timelines that used to take weeks or months down to days, while reaching larger and more diverse participant pools than a human-moderated study could staff. That piece is direct about the limit: this is "not a full replacement for human researchers." MIT Sloan Management Review makes a related case, describing generative AI compressing research timelines through synthetic "digital twin" testing, AI-moderated interviews, and large-scale analysis of open-ended responses — but framing the shift as AI folded into the process "with humans in the loop," not AI running the process by itself. Both pieces expand where qualitative insight can be applied. Neither replaces the researchers who decide what an insight means.
How is Sweetgreen using AI-moderated interviews to change its menu?
Sweetgreen's research team used to choose on every project: move fast with thin data, or move slow enough to get the "why" behind it. According to the company's customer story with Listen Labs, Sweetgreen no longer has to choose. Brian Davia, the company's Head of Consumer and Business Insights, describes getting "five times as many responses" for "half the cost," with results "turned around and analyzed in about a tenth of the time" it used to take.
That speed changed what got built. Sweetgreen's wraps went from tested concept to a 67-restaurant rollout, and its 100g protein bowl came directly out of interviews run this way, built "in direct response to customer demand." The interviews told Sweetgreen what customers wanted, at a scale that would have taken years the old way. Deciding to build the bowl, price it, and put it on the menu across roughly 300 locations was still a call Sweetgreen's team made. CEO Jonathan Neman put it plainly: "Ultimately we made the decision to do what the customer wants and figure it out."
What did Anthropic find when it studied how people use AI at work?
Anthropic built something different: an in-house system called Anthropic Interviewer, which the company used to run 1,250 structured interviews with professionals about how they use AI at work, split across a general workforce sample, working scientists, and creative professionals.
Human researchers show up at two points in that process. Before a single interview ran, "human researchers collaborated with Anthropic Interviewer to make any necessary edits to finalize the plan," the company's write-up of the study says. After the interviews were complete, a human researcher again worked with the system, this time to analyze the transcripts and turn them into answers to Anthropic's research questions.
The findings themselves sound like the kind of thing a research team wants to be confident about. 86% of participants said AI saves them time, and most kept the parts of their job central to their professional identity while handing off routine work. Anthropic's team didn't treat that as settled. The write-up states plainly, "we cannot determine whether AI usage directly caused these outcomes," and frames the whole study as producing "early signals" rather than conclusions. The company that built the AI interviewer, running this study, still chose to hedge the results rather than let scale stand as proof.
Why does Microsoft still route its research through a research lead?
Romani Patel, a senior research manager at Microsoft, has cut research timelines from four to six weeks down to results in days, sometimes hours. When her team collected customer stories about how people use Copilot for Microsoft's 50th-anniversary celebration, video interviews that would have taken six to eight weeks came back in a single day. Patel told VentureBeat why the speed matters: "By the time we get to them, either the decision has been made or we lose out on the opportunity to actually influence it."
Patel is describing speed as what makes a decision possible before the moment to make it has passed. The interviews still needed someone who could recognize a usable Copilot story when the transcripts came back, and decide which ones represented Microsoft. The speed simply meant that judgment happened in time to matter.
Where's the line between collecting research and deciding what it means?
Three companies, two ways of getting there, the same outcome every time. Sweetgreen and Microsoft bought a vendor platform. Anthropic built one in-house. Either way, interviews that used to cap out around a few dozen respondents now run into the hundreds, on a timeline measured in days instead of quarters, sometimes hours. None of the three moved the next step onto whichever system did the collecting, the one where somebody decides what the pattern means and commits real resources because of it.
Sweetgreen's Davia and Microsoft's Patel both want one thing from faster collecting — more time for the harder part, weighing a pattern against everything else the business knows before deciding what to build. Anthropic's team arrived at that conclusion from the other direction. Its write-up had the scale and the numbers, and its researchers still chose to call the results "early signals" instead of a verdict. Automating the interview makes sense, because the interview is a data-collection problem, and data-collection problems are the ones AI is good at solving today. Keeping the decision with a person makes sense too, because a decision requires weighing a finding against context the interview data alone can't carry, things like the roadmap, the brand, last quarter's numbers, and whatever a competitor just shipped.
That split isn't unique to interview platforms. Why your next analytics tool will create another silo, not solve the last one makes an identical argument about dashboards and BI tools. A tool that captures signals faster doesn't close the gap between data and a decision, because that gap sits between the tools rather than inside any single one. Run enough of those tools at once and a team runs into cross-source questions it still can't answer. The data isn't the missing piece. Nobody owns turning it into a decision.
A customer intelligence layer is supposed to connect what customers say to what a team decides to do about it. Collecting the "what customers say" half faster leaves the second half of that job exactly as large. If anything, that second half becomes the only place left where a research team's judgment shows up at all.
Caveats
Not every research-driven decision needs to stay this manual. Low-stakes, high-volume calls, like which of two subject lines gets a slightly higher open rate or which onboarding screen order tests marginally better, are already automated at a lot of product organizations, and reasonably so: the cost of being wrong is small and reversible. The argument here is narrower, about decisions that commit real resources or confidence on the basis of a qualitative pattern, the kind Sweetgreen made about a menu item, Microsoft made about which customer stories represent the company, or Anthropic made about how much weight to put on its study's findings. None of the three companies in this piece has published a policy stating they'll never let a model make that call. What they've published is what they've done so far: kept a person's judgment in the loop. That describes current practice. Next year's decisions may look different.
—Steven Rencher, Founder of Monadux
FAQ
What does AI automate in customer research today?
Collection. AI-moderated platforms handle recruiting participants, running adaptive interviews across video, voice, and text, and processing responses at a scale of hundreds rather than dozens. Anthropic, Microsoft, and Sweetgreen all use AI this way today. None of them use it to decide what the findings mean.
Can AI replace human judgment in customer research?
Not at the companies in this piece. Microsoft's Romani Patel and Sweetgreen's Brian Davia both describe a named person deciding what to build because of what AI collects. Anthropic's researchers describe a similar pattern from the other side. Their write-up had the scale and the numbers, and they still chose to call the results "early signals" instead of a verdict.
Which real companies use AI to run customer interviews at scale?
Sweetgreen and Microsoft both use Listen Labs' AI-moderated interview platform, according to each company's published customer story and independent reporting from TechCrunch and VentureBeat. Anthropic built a different tool instead, called Anthropic Interviewer, and used it to study how 1,250 professionals use AI at work.
What's the difference between a research finding and a research decision?
A finding is what customers said, or what an interview turned up, gathered and processed by the interview platform. A decision is what a company does about it, or how much confidence to place in it, like Sweetgreen building a new menu item or Anthropic choosing to call a large study's results "early signals" instead of settled conclusions. AI produces the finding. A person still makes the decision.
Does AI-moderated research remove the need for a research team?
No. Every company in this piece used AI to expand its research team's reach rather than replace it. Sweetgreen and Microsoft still employ the research leads who ran interviews before AI moderation existed, and Anthropic's research team still decided how to interpret the study its tool ran. The tool changed how much ground one team can cover. The role itself stayed exactly as necessary.